Shadow AI: how to find the agents already running in your company
Shadow AI is any AI agent or integration running against company systems that is not on an approved register. It appears when a team ships an integration without review, when an owner leaves and nobody inherits their service, or when a prototype quietly becomes production. You find it by comparing what your systems observe against what your register claims, not by sending a survey.
Why surveys do not work
The usual first attempt is to ask. This under-reports for structural reasons: people do not classify their script as an AI agent, the person who built it has moved teams, or the integration is genuinely forgotten because it has run without incident for a year.
Discovery has to be observational. The question is not "what do you run" but "what is actually calling our systems, and which of those things is on the register".
The five signals worth detecting
Most shadow AI resolves to a small number of detectable patterns:
- An agent identifier active in production with no matching registry entry.
- An agent whose registered owner has left the organisation, so nobody is accountable for it.
- An agent repeatedly reaching a resource it was never scoped for — often scope creep rather than attack.
- Credentials still valid long after the project that issued them ended.
- Production agents with no named owner at all.
Discovery is a governance problem, not a detection problem
The detection itself is not technically difficult. The hard part is what happens next: an unregistered agent found on Friday is only useful if someone owns the decision to register, scope or retire it by Monday.
This is why discovery works best attached to a register rather than as a standalone scan. A finding that links to a one-click "register this agent and assign an owner" is acted on. A finding in a PDF is not.
What to do with what you find
Resist the urge to switch everything off. Most shadow AI is doing legitimate work, and an abrupt shutdown teaches teams to route around governance next time. The durable sequence is: register it, name an owner, declare its intended scope, then observe whether it stays inside that scope.
Only once you know what a thing is supposed to do can you meaningfully say it has done something wrong.
Frequently asked
What is shadow AI?
Any AI system, agent or integration operating against company data or systems without being on an approved inventory. It is the AI equivalent of shadow IT, and it arises the same way: teams solving real problems faster than the approval process moves.
How do you detect unregistered AI agents?
By comparing observed activity against your register. Agent identifiers appearing in telemetry with no registry entry, ownerless production agents, stale credentials and out-of-scope resource access are the highest-yield signals.
Is shadow AI a security risk or a compliance risk?
Both, and the compliance exposure is usually the more immediate one. An unregistered agent cannot be risk-classified, cannot have oversight applied, and will not appear in any inventory you present to an auditor or regulator.
Bulwark does this in production
Agent registry, scope enforcement, human approvals and a hash-chained evidence trail you can verify without trusting us. Free Developer tier, 100,000 governed decisions a month, no card required.
Related guides
- AI agent inventory: the register regulators will ask forWhat belongs in an AI system inventory, why "a spreadsheet of models" is the wrong shape, and how to keep it accurate as agents change.
- EU AI Act Article 14: what human oversight actually requiresArticle 14 requires human oversight of high-risk AI systems. What that means in practice, what counts as evidence, and what auditors look for.
- Why blocking your AI agents is the wrong first moveEnforcement-first AI governance stalls because nobody will put an unproven gate in a production path. Monitor mode is how deployment actually happens.